China's AI-Powered Disinformation Targets Taiwan Elections

Prompt leak reveals AI-generated content targeting Taiwan

Researchers have documented concrete evidence of China deploying artificial intelligence to generate disinformation aimed at Taiwanese audiences. A research team led by Dr. Austin H. Wang, a political scientist at the University of Nevada, Las Vegas, and RAND, discovered an undeleted AI prompt in a post on a Chinese content farm. The prompt explicitly instructed writers to target a Taiwanese audience, rewrite content in Traditional Chinese while maintaining historical accuracy, and limit output to 500 words.

The line was removed within two minutes of publication but remained visible in the post's editing history. The incident has been logged in the OECD's AI Incidents and Hazards Monitor and covered by multiple Taiwanese news outlets, as first reported by AI Watch.

Coordinated networks combine AI with traditional tactics

This discovery fits within a broader pattern of sophisticated influence operations. Meta disrupted a China-originating network in March that used Taiwan-based proxy IPs to appear locally authentic. The network promoted pro-Beijing narratives and criticized Taiwan's ruling party across multiple Facebook pages, disguising itself as legitimate advertisers by paying $15,000 in ad fees using Hong Kong Dollars, Chinese Yuan, and Taiwan New Dollars.

Separate operations documented by OpenAI in February and June showed PRC-linked campaigns using large language models for both content creation and strategic planning. These efforts targeted Japanese Prime Minister Sanae Takaichi and US policy debates, demonstrating that AI-assisted influence operations span multiple countries.

During Taiwan's 2024 election, the Australian Strategic Policy Institute identified a campaign by Spamouflage (also called Dragonbridge), the CCP's largest network of inauthentic social media accounts. The operation deployed AI-generated news anchors to spread content from "The Secret History of Tsai Ing-wen," targeting the outgoing president. Google also documented this incident. Analysis showed the campaign "had limited reach, with practically no engagement from organic users."

Why it matters

The persistence of these campaigns despite low engagement suggests China's goal is not persuasion but distraction and erosion of trust. When fake content becomes easy to produce, people may grow unwilling to believe anything—a phenomenon called the "liar's dividend." This allows actors to dismiss genuine evidence as fabricated simply because fabrication is now commonplace. The threat lies not in individual fake posts but in the scale, speed, and coordination of operations that can overwhelm public discourse.

Policy responses lag behind operational sophistication

Current detection efforts from AI companies and social media platforms focus primarily on identifying and blocking bad actors rather than analyzing collaborative behavior patterns and operational intent. Meta and OpenAI have noted that focusing on behavior rather than content appears effective, but this approach has not become standard practice across the industry.

The European Union's Digital Services Act legally requires Very Large Online Platforms to assess and mitigate risks to public discourse. The proposed US Platform Accountability and Transparency Act would require large platforms to open data to independent researchers. Taiwan's National Communications Commission drafted a platform regulation bill in 2022, but it stalled over free speech concerns.

Ahead of Taiwan's 2026 elections, the District Prosecutors' Office has established a task force to combat deepfake content, and the Ministry of Digital Affairs plans to launch a reporting network in August. Taiwan participated in a Global Cooperation and Training Framework workshop with the US and partners in July addressing AI misuse and disinformation.

However, these mechanisms remain focused on identifying fake content itself rather than addressing how content is distributed, amplified, and coordinated across networks.

These details were first reported by AI Watch.